1. What is Accelerated Logic AI, actually?
Most AI chats lock you into one model and store everything on their cloud. Accelerated Logic AI is the opposite:
🔒 Local-first
Chat history, files, skills, and memories are stored in browser-managed IndexedDB and localStorage by default. Prompts and selected context go to any model provider you choose; some remote requests pass through the Accelerated Logic server proxy. The site also downloads assets needed for the app and local models.
⚡ Any Model
Puter.js or direct provider connections, local Ollama, and browser-local WebGPU models are available. Browser-local models need their runtime and model files downloaded first; only inference for that selected local model runs on your device. You can equip multiple models and switch between them.
🤖 Multi-agent + Sandboxes
Instead of one model guessing, you can spin up a team — Architect, Coder, Auditor, Checker — that collaborate, then render the result live in a Web Sandbox and execute Python right in the browser via WASM.
You don't need to learn all that on day one. This guide just gets you comfortable moving around.
2. The 4 Areas of the Workspace
A. Left Sidebar — Your Library
This is your file cabinet. It collapses on mobile for maximum writing space.
- Search bar at top: Searches both chat titles AND inside all messages. Try typing
python sortingto find past conversations instantly. - AI Platform section:
- Memory: Your long-term facts. Click it to inspect what the AI remembers about you.
- Skills: Your custom instructions. Shows active skills.
- Chats section:
- Top actions:
Folder+icon creates a new folder.+icon starts a New Chat. - Pinned Chats: Any chat you pin appears here first, always on top.
- Folders: Click a folder name to expand or collapse. The number badge shows chat count. If it has a brain icon, that folder has its own Folder Memory — custom instructions applied only to chats inside it.
- Other / All Chats: Uncategorized chats. Drag any chat card onto a folder to move it, or drop onto "Other Chats" header to un-folder it.
- Actions & Quick Rename: Hovering a chat card reveals Rename, Pin/Unpin, Export Markdown, and Delete. Double-click any chat name to rename inline.
- Top actions:
- Files section: Every file you upload lives here. Images show an image icon, documents show a file icon. Drag a file directly onto the composer to stage it for one prompt. Click the Bookmark icon to Pin as persistent context — attaching that file to every future prompt in that chat.
- Experimental section: Model Playground (test models side-by-side), Python Sandbox, Web Sandbox, Browser Agent Sync (shows green dot when active), and GitHub Integration.
B. Top Header — Controls
- Logo click: Navigate home to the main landing page.
- Collapse arrow: Hide the left sidebar to maximize editing workspace.
- Model Dropdown: Displays your active model (e.g. DeepSeek V4 Flash). Click to hot-swap equipped models mid-conversation or open Model Organizer.
- Settings, Model Organizer, Memory & Drive status: Drive shows
Syncing... / Saved / Errorand connection state. All configuration resides cleanly inside the settings modal.
C. Center Chat Canvas
User prompts appear on the right, assistant responses on the left.
- Assistant tools: Image previews, saved agent activity and scratchpad notes when a workflow provides them, and clean markdown with code-block copying. These records do not reveal a provider's private reasoning.
- Action Bar Below Messages: Regenerate (re-run answer), Model Switcher (run that prompt with a different model), Read Aloud (TTS audio with soundwave animation), Copy, Delete, Fork (branch conversation from that message), and live TPS (Tokens/Sec) + Context % badges.
D. Bottom Composer — Where You Type
- Plus icon (+): Upload files via picker (supports text, images, PDFs, videos, and ZIP archives which auto-extract).
- / Slash Command: Type
/to open the skill & persona selector. Filter skills (e.g./code-review) and press Enter to toggle. Equipped skills appear as active chips above the composer. - Microphone: Real-time voice dictation transcribed directly into the input.
- Wand (Enhance Prompt): Transforms brief inputs (e.g. "landing page") into structured, detailed prompts automatically.
- Send / Stop / Queue Response: Send transforms into a Stop square during generation + a Queue Response button to stack follow-up prompts.
- Global Drag & Drop: Drag any file from your computer anywhere over the app window to stage it for your message.
3. Your First 5 Minutes — Do This Now
Click New Chat in the sidebar. The title starts as "New Chat" and auto-updates based on your first prompt.
Type: Explain how useState works in React with a simple counter example. Observe streaming text and header trace logs.
Hover under your prompt: Click Edit and change to Now do it with TypeScript and Tailwind styled button. Notice it re-runs. Under assistant message, try Read Aloud or Regenerate.
Drag a code file into composer and send Review this file for bugs. In sidebar Files, click its Bookmark icon to Pin it as persistent context for all future prompts in this chat.
Click Folder+ -> Name: React Learning, set Folder Memory: Always explain with TypeScript examples and include accessibility notes. Drag your chat inside.
4. How Your Data Actually Works (Important)
- Local-First Privacy: Everything is stored locally by default. Clearing browser cache clears history unless exported. Use Export Markdown for backups.
- Google Drive Integration: Opt-in cloud sync uses Google's isolated AppData folder. It remains hidden from standard Drive view and sandboxed for Accelerated Logic. You can select individual sync toggles in Settings -> Google Drive.
- Context Isolation: AI models only receive files staged in the composer or pinned to the current chat. Memories are injected dynamically only when relevant.
5. Prompting Tips That Actually Matter Here
Use explicit instructions like Use React + Tailwind, no external deps rather than vague requests like "build app".
Include phrases like Generate and preview in canvas to trigger interactive live sandbox rendering.
For math, data analysis, or algorithms, ask to Calculate and verify with python to execute code via in-browser WASM Python sandbox.
End prompt with explicit request like Generate an image of... to trigger visual generation tools.
6. When You Get Stuck
Click the model dropdown or open Model Organizer to select or equip a model (such as free Puter.js models).
Ensure Ollama is running with CORS allowed. Run OLLAMA_ORIGINS="*" ollama serve in terminal.
Inside sandboxed iframe previews, browser security restricts media APIs. Click Open in New Tab in the top-right header to grant permissions.
Check the Context % badge in the assistant response bar. If >90%, start a New Chat to clear token overhead.
7. A Real Day-One Workflow: From Blank Chat to Shipped Component
Reading the sections above in isolation is useful; watching one complete loop is better. Here is a realistic first session, roughly ten minutes start to finish:
- Open a New Chat from the sidebar
+icon. The thread is titled automatically from your first prompt. - Pick a model. With no setup, the workspace defaults to AL Router · Fast, a free built-in route that needs no account (see Part 2). Or equip a free Puter.js model in the Model Organizer.
- Stage your context. Drag
types.tsfrom the sidebar Files section onto the composer, or press Ctrl/Cmd+V to paste a screenshot of the UI you want to imitate. - Prompt with intent: "Using the types in the attached file, build a typed StatsCard component in React + Tailwind. Preview in canvas." Asking for a canvas preview forces an interactive render inside the chat instead of a static code block (see Part 19).
- Iterate surgically. Follow up with "Add a delta percentage with green/red coloring and an aria-label". The preview updates in place — no need to restate the whole requirement each turn.
- Fork before you explore. Hover your last message, click Fork, and try a radically different direction in the branch while the working version stays untouched (see Part 8).
- Move to production. Click Open in Web Sandbox on the preview card, keep editing in the multi-file workspace, then Download ZIP or push to GitHub (Part 17).
- Tidy up. Pin the final component to the chat context if you will keep iterating, or unpin it when you are done so it stops consuming context tokens.
If any step misfires — white preview, wrong library, lost context — the troubleshooting sections in Parts 9, 19 and 20 cover the exact failure modes.
8. Where Accelerated Logic AI Runs Well — and Where It Doesn't
| Environment | What works | What to know |
|---|---|---|
| Chrome / Edge (desktop) | Everything: cloud models, WebLLM/WebGPU local models, on-device Whisper transcription, Kokoro voice, WASM Python | Best-supported environment. Keep hardware acceleration on (chrome://gpu). |
| Firefox (desktop) | Cloud chat, Puter.js, agents, Web Sandbox, Python (WASM) | WebGPU support varies by version; on-device model paths may fall back to WASM or be unavailable. |
| Safari (desktop / iOS) | Cloud chat, most editing features | WebGPU is limited or absent; on-device Whisper uses the WASM fallback; Dictation uses Web Speech APIs which may be restricted. |
| Mobile browsers | Cloud chat, reading history, quick prompts, voice dictation | Multi-agent runs and local model downloads are heavy; the sidebar collapses by default for writing space. |
| Installed PWA | Same app in a standalone window (the site ships a web app manifest with display: standalone) |
Install from the browser menu. Handy for a dedicated workspace window; storage is still per-browser-profile. |
9. What Survives a Cache Clear — and What Doesn't
"Clear browsing data" feels routine, but in a local-first app it is also a delete button for your workspace. Know exactly what lives where before you clear anything:
| Stored item | Lives in | Survives cache clear? |
|---|---|---|
| Chat history, files, folders, memories, skills, custom agents, MCP servers | IndexedDB | No — removed when site data is cleared |
| Settings (theme, model equipment, fallback order, tool permissions) | IndexedDB + localStorage mirror | No |
| Downloaded on-device model weights (WebLLM, Whisper, Kokoro) | Browser Cache API / storage | No — must be re-downloaded |
| Google Drive connection token | sessionStorage | Cleared when the tab closes or session storage is wiped |
| GitHub OAuth / PAT credentials | Credential store (session + local, see Part 22) | Session tier dies with the session; persistent tier dies with site data |
| Your Google Drive backup files | Your Google account (hidden AppData folder) | Yes — independent of this browser |
| Exported snapshot JSON on your disk | Wherever you saved it | Yes — this is the safety net |
Rule of thumb: before clearing site data or switching machines, export a snapshot from Settings → Local Backup & Export (or confirm Drive sync is current). Restoring from a snapshot takes minutes; re-creating three months of chats takes forever.
10. Twenty Pointer Moves Worth Learning Early
Enter sends; Shift+Enter inserts a newline. Type / to open the skill & persona menu, and filter by name (e.g. /secur).
Double-click any chat title in the sidebar to rename it inline — no menus required.
Drag a file from your OS anywhere over the app window to stage it; the drop target is the whole workspace, not just the composer.
The Bookmark icon on a sidebar file pins it to every prompt in this chat. Click again to unpin when you are done.
Rename, Pin, Export Markdown, Delete, Fork, Regenerate, Read Aloud, and the model switcher all live on hover — moving your mouse away hides them again.
The sidebar search matches message bodies, not just titles — try searching an error string you remember from weeks ago.
Escape closes modals, menus and the Help Center in order of focus — useful when you have three panels open and clicked the wrong one.
The Pause/Break key pauses an active agent run, and the Stop square in the composer halts streaming immediately.
Inside the GitHub workspace editor, Ctrl/Cmd+P opens the quick-open command palette to jump between files.
Forking from any message creates an independent branch; deleting or mangling a branch never touches the original thread.